🔬 mesoscale physics

Learning Multiscale Dynamics of Nanoparticle Self-Assembly

This paper presents a physics-informed, time-varying autoregressive framework that extracts interpretable partial differential equations and fundamental physical coefficients directly from liquid-phase transmission electron microscopy videos to model and predict the multiscale dynamics of nanoparticle self-assembly.

Original authors: Shuaifeng Li, Puquan Pan, Jonas Hallstrom, Nicholas Kotov, Paul Bogdan, Qian Chen, Xiaoming Mao

Published 2026-10-28

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Shuaifeng Li, Puquan Pan, Jonas Hallstrom, Nicholas Kotov, Paul Bogdan, Qian Chen, Xiaoming Mao

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where tiny building blocks, far smaller than a grain of sand, can arrange themselves into complex, useful structures without anyone telling them exactly where to go. This is the promise of nanoparticle self-assembly, a process nature uses to create everything from the shells of microscopic organisms to the intricate patterns in butterfly wings. Scientists have long known that these tiny particles, when suspended in a liquid, are pulled together by invisible forces like electrical attraction and repulsion, eventually forming ordered crystals or networks. However, watching this happen has always been like trying to understand a conversation by listening to a single, chaotic shout. The particles move in a jittery, unpredictable way, influenced by their immediate neighbors and the fluid around them, making it nearly impossible to write down a simple set of rules that explains how they move from a scattered mess into a perfect, organized structure.

Until now, the gap between watching these particles and truly understanding the physics behind their movement has been a major hurdle. Researchers could film the process with powerful microscopes, capturing the motion of individual particles in real time, but turning those thousands of frames of jittery movement into a clear, predictive law of nature remained out of reach. The challenge was not just the sheer amount of data, but the fact that the rules governing the particles change as the system evolves. What drives a particle to move when it is alone in a fluid is different from what drives it when it is part of a growing crystal cluster. To solve this, scientists needed a way to listen to the collective voice of the particles rather than just tracking their individual, noisy steps.

A team of researchers has now developed a new method to decode these complex movements, turning chaotic video footage into a clear set of physical laws. They focused on gold nanocubes, tiny particles shaped like cubes with slightly cut corners, measuring about 63 nanometers across. When these particles are placed in a liquid cell and observed under a liquid-phase transmission electron microscope, they begin to drift and collide. As the experiment progresses, the particles stop moving randomly and start locking together into larger, crystalline networks. The researchers recorded this entire process, capturing the position and orientation of every single particle thousands of times per second.

Instead of trying to predict the path of every single particle, the team took a different approach. They treated the entire field of view as a continuous landscape, dividing it into a grid of small patches. For each patch, they calculated two main things: how many particles were present (the density) and how they were aligned (the orientation). This transformed the messy, discrete dots of individual particles into smooth, flowing fields, much like turning a cloud of individual raindrops into a continuous sheet of water. They then used a mathematical framework to learn how these fields change from one moment to the next. The key insight was to assume that what happens in any small patch depends only on what is happening in that patch and its immediate neighbors, a rule that reflects the physical reality that particles cannot instantly affect distant parts of the system.

By training this system on the video data, the researchers discovered a sequence of simple, interpretable rules that could accurately predict the future state of the assembly. These rules were not just a black-box prediction; they were structured in a way that revealed the underlying physics. When the team analyzed these learned rules over time, they found a clear signal that marked the exact moment the system switched from a disordered, fluid-like state to an ordered, solid-like state. This transition happened at a specific point in the experiment, around 110 seconds, a finding that matched independent measurements of the system's structure. The method successfully identified that the long-term evolution of the entire assembly is driven by just three dominant patterns: a general drive for the particles to come together, and two competing patterns that determine how different crystal regions grow and push against each other.

Digging deeper, the researchers examined how a small disturbance in one part of the system spreads to the rest. They found that when a local cluster of particles was nudged, the effect rippled outward in a way that matched the behavior of diffusion, where information spreads slowly and steadily through a medium, rather than shooting out like a bullet. This confirmed that the particles are moving in a highly viscous, liquid environment where their motion is dominated by friction and random collisions. Furthermore, by breaking down the complex interaction rules into their simplest components, the team identified specific mechanisms at play. One mechanism showed that as particles crowd together, they actually struggle to align perfectly, creating a tension between density and order. Another mechanism revealed that once a critical density is reached, local crowding suddenly triggers the particles to snap into alignment, driving the formation of crystal grains.

Perhaps most significantly, the team was able to translate these learned rules back into the language of classical physics, extracting the specific equations that govern the system. They found that the movement of the particles follows a set of coupled equations that describe how density and orientation change over space and time. These equations naturally included terms for diffusion, where particles spread out, and reaction terms, where local interactions stabilize the structure. Remarkably, the researchers did not force these equations to look a certain way; the symmetry and structure emerged naturally from the data. This suggests that the self-assembly process is governed by fundamental principles of thermodynamics, operating near a state of equilibrium even as it builds complex structures.

The study also highlighted that not all self-assembly is the same. In some experiments and computer simulations, the particles aligned gradually and smoothly. In the specific gold nanocube experiment, however, the alignment happened abruptly, driven by a unique instability where the orientation of the particles became unstable and sharpened rapidly. The new method was able to distinguish between these different pathways, identifying that the sign of a specific physical coefficient determined whether the system would order slowly or snap into place quickly. This ability to extract the precise physical coefficients from raw video data means that scientists can now move beyond simply observing these phenomena to understanding the exact levers that control them.

This work offers a powerful new tool for the future of materials science. By converting noisy, high-speed videos into clear, interpretable physical laws, the researchers have provided a way to see the hidden rules that govern how tiny building blocks organize themselves. This approach does not just describe what happened in a single experiment; it provides a framework that can be applied to other systems, from biological pattern formation to the manufacturing of new electronic materials. The ability to predict how a system will evolve and to identify the specific physical mechanisms driving that evolution opens the door to designing self-assembling structures with precision. Instead of hoping that particles will find the right arrangement, scientists may soon be able to steer the process, adjusting conditions in real time to guide the assembly toward a desired outcome. The chaotic dance of nanoparticles is no longer a mystery; it is a language that has finally been translated.

Technical Summary: Learning Multiscale Dynamics of Nanoparticle Self-Assembly

Problem Statement
Nanoparticle self-assembly is a critical pathway for creating programmable functional materials, yet its underlying dynamics remain difficult to decipher. These dynamics arise from multiscale, many-body, and environment-dependent interactions (van der Waals, electrostatic, ligand-mediated, solvent effects) that result in stochastic rearrangements, kinetic arrest, and collective reorganization. While Liquid-Phase Transmission Electron Microscopy (LPTEM) enables real-time observation of these processes with nanometer spatial and millisecond temporal resolution, extracting explicit physical laws from the resulting highly stochastic trajectories is a significant challenge. Conventional analytical approaches, such as Langevin equations, often lack closed forms or are computationally prohibitive in heterogeneous environments. Furthermore, while machine learning (e.g., deep learning) has advanced particle tracking and potential learning, these methods typically yield implicit representations that do not directly reveal explicit physical laws or governing equations.

Methodology
The authors develop a physics-informed, time-varying autoregressive (TVAR) framework to learn interpretable local dynamical operators directly from LPTEM videos. The methodology proceeds through the following stages:

Coarse-Graining to Eulerian Fields: Discrete, stochastic particle trajectories are mapped onto a regular spatial grid ( $N_x \times N_y$ $N_{x} \times N_{y}$ ) to generate continuous Eulerian fields. For corner-truncated gold nanocubes, three fields are defined:
- Local density ( $\rho$ ).
- Tetratic orientational order parameters ( $Q_1 = \text{Re}\langle e^{i4\theta} \rangle$ and $Q_2 = \text{Im}\langle e^{i4\theta} \rangle$ ), derived from the four-fold rotational symmetry of the cubes.
Time-Varying Autoregressive Modeling: The system evolution is modeled as $X_{t+1} = A_t X_t$ , where $X_t = [\rho_t, Q_{1,t}, Q_{2,t}]^T$ . The operator $A_t$ is a $3N \times 3N$ block matrix (where $N$ is the number of spatial patches) representing a sequence of linear operators that advance the state in time.
Constrained Optimization: To ensure physical interpretability and prevent overfitting, the optimization of $A_t$ $A_{t}$ is constrained to a sparse, nearest-neighbor manifold. The loss function $L_t$ $L_{t}$ balances data fidelity with three regularization terms:
- Frobenius norm penalty ( $\lambda_1$ ) for numerical stability.
- Temporal smoothness penalty ( $\lambda_2$ ) to prevent abrupt operator changes.
- Graph Laplacian penalty ( $\lambda_3$ ) to enforce spatial consistency (neighboring patches obey similar rules).
- Hard Structural Prior: Each block of $A_t$ is restricted to a sparse, nearest-neighbor banded structure, encoding the physical principle that local interactions dominate.
Optimization Algorithm: The constrained optimization is solved at each time step using Riemannian gradient descent on the manifold of sparse matrices.
Multi-Scale Analysis:
- Macroscopic: Operator volatility ( $\|A_{t+1} - A_t\|_F$ ) tracks phase transitions. Singular Value Decomposition (SVD) of the cumulative propagator $\Phi = \prod A_t$ identifies dominant structural modes.
- Mesoscopic: Discrete Green's functions are derived from partial propagators to analyze impulse responses and diffusion scaling.
- Microscopic/Continuum: Canonical Polyadic (CP) decomposition of the interaction tensor isolates distinct coupling mechanisms. Finally, spatial moments of the learned stencils are used to extract coefficients for continuum Partial Differential Equations (PDEs).

Key Results
The framework was applied to two LPTEM experiments of gold nanocube self-assembly and a Molecular Dynamics (MD) simulation.

Reconstruction Accuracy: The sparse operators successfully reconstructed the full assembly dynamics with a cumulative relative error of approximately 1.4%, confirming the capture of dominant spatiotemporal dynamics.
Phase Transition Detection: Operator volatility revealed a sharp peak at approximately 110 seconds, corresponding to the transition from a mobile disordered phase to an immobile ordered crystalline phase. This timing was independently corroborated by Ollivier–Ricci curvature analysis of particle networks.
Low-Rank Collective Dynamics: SVD of the cumulative propagator showed that despite complex raw trajectories, the long-timescale dynamics are dominated by only three modes: a monopole (uniform crystallization drive) and two dipoles (symmetry-breaking gradients driving grain boundary formation).
Diffusive Transport: Green's function analysis demonstrated that perturbations propagate diffusively with a scaling of $R \sim t^{0.5}$ , consistent with dissipative dynamics in a viscous liquid environment, rather than ballistic transport.
Interaction Mechanisms: CP decomposition identified three distinct interaction components:
1. Global density accumulation where crowding suppresses orientational order.
2. Spatially localized crystallization where density drives alignment (active during the phase transition).
3. A mechanism involving density-to-orientation coupling and orientational redistribution, which reverses its physical role (from weakening to reinforcing alignment) across the phase transition.
Emergent Symmetries: The learned interaction stencils exhibited approximate reciprocity (symmetry between coupling $\alpha \to \beta$ and $\beta \to \alpha$ ), a manifestation of Onsager reciprocal relations, despite not being explicitly enforced in the algorithm.
PDE Extraction: The method successfully extracted coefficients for a coupled reaction-advection-diffusion system. Notably, the orientational self-diffusion coefficients were found to be negative in the primary experiment (indicating anti-diffusive instability driving domain sharpening), contrasting with positive values in the MD simulation and a second experiment. This distinction allowed the framework to differentiate between abrupt ordering and gradual relaxation pathways.

Significance and Claims
The paper claims to provide a robust, computationally efficient tool for discovering hidden physical mechanisms from observational data without relying on black-box models. By bridging discrete experimental trajectories, interpretable data-driven operators, and continuum field theories, the framework offers a unified scope to:

Extract simple organizing principles underlying complex, multiscale self-assembly.
Directly infer fundamental coefficients (diffusion, advection, reaction) and identify the governing PDEs.
Distinguish between different kinetic pathways (e.g., anti-diffusive instability vs. standard diffusion) based on the sign of extracted coefficients.
Enable the potential for predictive design of self-assembled nanostructures by steering assembly toward desired outcomes based on real-time transport coefficient estimates.

The authors emphasize that while the current framework assumes linearity at each time step and nearest-neighbor interactions, it successfully recovers physical laws (including Onsager reciprocity and Ginzburg-Landau-like structures) that emerge naturally from the data, offering a path toward interpretable physics discovery in complex many-body systems.

Technical Summary: Learning Multiscale Dynamics of Nanoparticle Self-Assembly

More like this